1: \begin{abstract}

2: Based on the joint bidiagonalization process of a large matrix pair $\{A,L\}$,

3: we propose and develop an iterative regularization algorithm for the large scale

4: linear discrete ill-posed problems in general-form regularization:

5: $\min\|Lx\| \ \mbox{{\rm subject to}} \ x\in\mathcal{S} =

6: \{x|\ \|Ax-b\|\leq \tau\|e\|\}$ with a Gaussian white noise $e$

7: and $\tau>1$ slightly, where $L$ is a regularization matrix.

8: Our algorithm is different from the hybrid one proposed by Kilmer {\em et al.},

9: which is based on the same process but solves

10: the general-form Tikhonov regularization problem:

11: $\min_x\left\{\|Ax-b\|^2+\lambda^2\|Lx\|^2\right\}$.

12: We prove that the iterates take the

13: form of attractive filtered generalized singular value decomposition (GSVD)

14: expansions, where the filters are given explicitly. This result and

15: the analysis on it show that the method must have the desired semi-convergence property

16: and get insight into the regularizing effects of the method.

17: We use the L-curve criterion or

18: the discrepancy principle to determine $k^*$.

19: The algorithm is simple and effective, and numerical experiments illustrate

20: that it often computes

21: more accurate regularized solutions than the hybrid one.

22: \end{abstract}